Automatic Synonym Discovery with Knowledge Bases
Automatic Synonym Discovery with Knowledge Bases
arXiv:1706.08186v1 [cs.CL] 25 Jun 2017
Meng
University of Illinois at Urbana-Champaign mengqu2@illinois.edu
Xiang Ren
University of Illinois at Urbana-Champaign
xren7@illinois.edu
Jiawei Han
University of Illinois at Urbana-Champaign
hanj@illinois.edu
ABSTRACT
Recognizing entity synonyms from text has become a crucial task in many entity-leveraging applications. However, discovering entity synonyms from domain-speci c text corpora (e.g., news articles, scienti c papers) is rather challenging. Current systems take an entity name string as input to nd out other names that are synonymous, ignoring the fact that o en times a name string can refer to multiple entities (e.g., "apple" could refer to both Apple Inc and the fruit apple). Moreover, most existing methods require training data manually created by domain experts to construct supervisedlearning systems. In this paper, we study the problem of automatic synonym discovery with knowledge bases, that is, identifying synonyms for knowledge base entities in a given domain-speci c corpus.
e manually-curated synonyms for each entity stored in a knowledge base not only form a set of name strings to disambiguate the meaning for each other, but also can serve as "distant" supervision to help determine important features for the task. We propose a novel framework, called DPE, to integrate two kinds of mutuallycomplementing signals for synonym discovery, i.e., distributional features based on corpus-level statistics and textual pa erns based on local contexts. In particular, DPE jointly optimizes the two kinds of signals in conjunction with distant supervision, so that they can mutually enhance each other in the training stage. At the inference stage, both signals will be utilized to discover synonyms for the given entities. Experimental results prove the e ectiveness of the proposed framework.
ACM Reference format: Meng , Xiang Ren, and Jiawei Han. 2017. Automatic Synonym Discovery with Knowledge Bases. In Proceedings of KDD'17, August 13-17, 2017, Halifax, NS, Canada, , 9 pages. DOI: 10.1145/3097983.3098185
1 INTRODUCTION
People o en have a variety of ways to refer to the same real-world entity, forming di erent synonyms for the entity (e.g., entity United States can be referred using "America" and "USA"). Automatic synonym discovery is an important task in text analysis and understanding, as the extracted synonyms (i.e. the alternative ways to refer to the same entity) can bene t many downstream applications [1, 32, 33, 37]. For example, by forcing synonyms of an entity to be assigned in the same topic category, one can constrain the topic modeling process and yield topic representations with higher quality [33]. Another example is in document retrieval [26], where
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for pro t or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permi ed. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior speci c permission and/or a fee. Request permissions from permissions@. KDD'17, August 13-17, 2017, Halifax, NS, Canada ? 2017 ACM. 978-1-4503-4887-4/17/08. . . $15.00 DOI: 10.1145/3097983.3098185
Text Corpus
ID
Sentence
1 Washington is
a
s tate
i n
the
Pacific
N orthwest
region.
2 Washington served
as
the
f irst
President
o f
t he
US.
3 The
exact
c ause
of
leukemia is
unknown. 4 Cancer involves
a bnormal
c ell
g rowth.
Synonym Seeds
Cancer Cancer,
C ancers
Washington
S tate Washington
State,
State
of
Washington
Washington State
Cancer
George Washington Leukemia
Entities
Knowledge Bases
Figure 1: Distant supervision for synonym discovery. We link entity mentions in text corpus to knowledge base entities, and collect training seeds from knowledge bases.
we can leverage entity synonyms to enhance the process of query expansion, and thus improve the retrieval performances.
One straightforward approach for obtaining entity synonyms is to leverage publicly available knowledge bases such as Freebase and WordNet, in which popular synonyms for the entities are manually curated by human crowds. However, the coverage of knowledge bases can be rather limited, especially on some newly emerging entities, as the manual curation process entails high costs and is not scalable. For example, the entities in Freebase have only 1.1 synonyms on average. To increase the synonym coverage, we expect to automatically extract more synonyms that are not in knowledge bases from massive, domain-speci c text corpora. Many approaches address this problem through supervised [19, 27, 29] or weakly supervised learning [11, 20], which treat some manually labeled synonyms as seeds to train a synonym classi er or detect some local pa erns for synonym discovery. ough quite e ective in practice, such approaches still rely on careful seed selections by humans.
To retrieve training seeds automatically, recently there is a growing interest in the distant supervision strategy, which aims to automatically collect training seeds from existing knowledge bases. e typical work ow is: i) detect entity mentions from the given corpus, ii) map the detected entity mentions to the entities in a given knowledge base, iii) collect training seeds from the knowledge base. Such techniques have been proved e ective in a variety of applications, such as relation extraction [10], entity typing [17] and emotion classi cation [14]. Inspired by such strategy, a promising direction for automatic synonym discovery could be collecting training seeds (i.e., a set of synonymous strings) from knowledge bases.
Although distant supervision helps collect training seeds automatically, it also poses a challenge due to the string ambiguity problem, that is, the same entity surface strings can be mapped to di erent entities in knowledge bases. For example, consider the string "Washington" in Figure 1. e "Washington" in the rst sentence represents a state of the United States; while in the second sentence it refers to a person. As some strings like "Washington"
have ambiguous meanings, directly inferring synonyms for such strings may lead to a set of synonyms for multiple entities. For example, the synonyms of entity Washington returned by current systems may contain both the state names and person names, which is not desirable. To address the challenge, instead of using ambiguous strings as queries, a be er way is using some speci c concepts as queries to disambiguate, such as entities in knowledge bases.
is motivated us to de ne a new task: automatic synonym discovery for entities with knowledge bases. Given a domain-speci c corpus, we aim to collect existing name strings of entities from knowledge bases as seeds. For each query entity, the existing name strings of that entity can disambiguate the meaning for each other, and we will let them vote to decide whether a given candidate string is a synonym of the query entity. Based on that, the key task for this problem is to predict whether a pair of strings are synonymous or not. For this task, the collected seeds can serve as supervision to help determine the important features. However, as the synonym seeds from knowledge bases are usually quite limited, how to use them e ectively becomes a major challenge. ere are broadly two kinds of e orts towards exploiting a limited number of seed examples.
e distributional based approaches [9, 13, 19, 27, 29] consider the corpus-level statistics, and they assume strings which o en appear in similar contexts are likely to be synonyms. For example, the strings "USA" and "United States" are usually mentioned in similar contexts, and they are the synonyms of the country USA. Based on the assumption, the distributional based approaches usually represent strings with their distributional features, and treat the synonym seeds as labels to train a classi er, which predicts whether a given pair of strings are synonymous or not. Since most synonymous strings will appear in similar contexts, such approaches usually have high recall. However, such strategy also brings some noise, since some non-synonymous strings may also share similar contexts, such as "USA" and "Canada", which could be labeled as synonyms incorrectly.
Alternatively, the pa ern based approaches [5, 15, 20, 22] consider the local contexts, and they infer the relation of two strings by analyzing sentences mentioning both of them. For example, from the sentence " e United States of America is commonly referred to as America.", we can infer that "United States of America" and "America" have the synonym relation; while the sentence " e USA is adjacent to Canada" may imply that "USA" and "Canada" are not synonymous. To leverage this observation, the pa ern based approaches will extract some textual pa erns from sentences in which two synonymous strings co-occur, and discover more synonyms with the learned pa erns. Di erent from the distributional based approaches, the pa ern based approaches can treat the pa erns as concrete evidences to support the discovered synonyms, which are more convincing and interpretable. However, as many synonymous strings will not be co-mentioned in any sentences, such approaches usually su er from low recall.
Ideally, we would wish to combine the merits of both approaches, and in this paper we propose such a solution named DPE (distributional and pa ern integrated embedding framework). Our framework consists of a distributional module and a pa ern module. e distributional module predicts the synonym relation from the global distributional features of strings; while in the pa ern module, we aim to discover synonyms from the local contexts. Both modules are built on top of some string embeddings, which preserve the
Target
S trings America
USA
ID
Sentence
1 The USA
is
also
k nown
a s America
.
2 The USA
( America
) is
a
country
o f
50
states .
3 The
USA is
a
highly
d eveloped
c ountry
.
4 Canada is
a djacent
t o America
.
Pattern Based Approaches Local Patterns:
USA ; known as ; America USA ; ( ) ; America
Distributional Based Approaches
Country State The
America 1
1 2
USA
2
1 2
Figure 2: Comparison of the distributional based and pattern based approaches. To predict the relation of two target strings, the distributional based approaches will analyze their distributional features, while the pattern based approaches will analyze the local patterns extracted from sentences mentioning both of the target strings.
semantic meanings of strings. During training, both modules will treat the embeddings as features for synonym prediction, and in turn update the embeddings based on the supervision from synonym seeds. e string embeddings are shared across the modules, and therefore each module can leverage the knowledge discovered by the other module to improve the learning process.
To discover missing synonyms for an entity, one may directly rank all candidate strings with both modules. However, such strategy can have high time costs, as the pa ern module needs to extract and analyze all sentences mentioning a pair of given strings when predicting their relation. To speed up synonym discoveries, our framework will rst utilize the distributional module to rank all candidate strings, and extract a set of top ranked candidates as high-potential ones. A er that, we will re-rank the high-potential candidates with both modules, and treat the top ranked candidates as the discovered synonyms.
e major contributions of the paper are summarized as follows: ? We propose to study the problem of automatic synonym discovery
with knowledge bases, i.e., aiming to discover missing synonyms for entities by collecting training seeds from knowledge bases. ? We propose a novel approach DPE, which naturally integrates the distributional based approaches and the pa ern based approaches for synonym discovery. ? We conduct extensive experiments on the real world text corpora. Experimental results prove the e ectiveness of our proposed approach over many competitive baseline approaches.
2 PROBLEM DEFINITION
In this section, we de ne several concepts and our problem: Synonym. A synonym is a string (i.e., word or phrase) that means exactly or nearly the same as another string in the same language [21]. Synonyms widely exist in human languages. For example, "Aspirin" and "Acetylsalicylic Acid" refer to the same drug; "United States" and "USA" represent the same country. All these pairs of strings are synonymous. Entity Synonym. For an entity, its synonym refers to strings that can be used as alternative names to describe that entity. For example, both the strings "USA" and "United States" serve as alternative names of the entity United States, and therefore they are the synonyms of this entity. Knowledge Base. A knowledge base consists of some manually constructed facts about a set of entities. In this paper, we only focus
Seed Collection
Text Corpus
ID
Sentence
1 Washington is
a
s tate
i n
the
Pacific
N orthwest
region.
2 Washington served
as
the
f irst
President
o f
t he
US.
3 The
exact
c ause
of
leukemia is
unknown.
4 Cancer involves
a bnormal
c ell
g rowth.
Knowledge Base
Synonym Seeds
Entity
Synonyms
Cancer
cancer, cancers
USA USA,
America,
U.S.A.
Illinois
Illinois,
I L
Model Learning
String Embeddings
Washington Washington
leukemia Cancer
US ......
......
Distributional Module Distributional Score Function
Pattern Module Pattern Classifier
Seeds Seeds
Figure 3: Framework Overview.
Inference
Query Entity
Discovered Synonyms
Distributional
High-Potential Candidates
Distributional
Pattern
on the existing entity synonyms provided by knowledge bases, and we will collect those existing synonyms as training seeds to help discover other missing synonyms.
Problem Definition. Given the above concepts, we formally de ne our problem as
follows.
De nition 2.1. (Problem De nition) Given a knowledge base K and a text corpus D, our problem aims to extract the missing synonyms for the query entities.
3 FRAMEWORK
In this section, we introduce our approach DPE for entity synonym discovery with knowledge bases. To infer the synonyms of a query entity, we leverage its name strings collected from knowledge bases to disambiguate the meaning for each other, and let them vote to decide whether a given candidate string is a synonym of the query entity. erefore, the key task for this problem is to predict whether a pair of strings are synonymous or not. For this task, the synonym seeds collected from knowledge bases can serve as supervision to guide the learning process. However, as the number of synonym seeds is usually small, how to leverage them e ectively is quite challenging. Existing approaches either train a synonym classi er with the distributional features, or learn some textual pa erns for synonym discovery, which cannot exploit the seeds su ciently.
To address this challenge, our framework naturally integrates the distributional based approaches and the pa ern based approaches. Speci cally, our framework consists of a distributional module and a pa ern module. Given a pair of target strings, the distributional module predicts the synonym relation from the global distributional features of each string; while the pa ern module considers the local contexts mentioning both target strings. During training, both modules will mutually enhance each other. At the inference stage, we will leverage both modules to nd high-quality synonyms for the query entities.
Framework Overview. e overall framework of DPE (Figure 3) is summarized below: (1) Detect entity mentions in the given text corpus and link them
to entities in the given knowledge base. Collect synonym seeds from knowledge bases as supervision. (2) Jointly optimize the distributional and the pa ern modules.
e distributional module predicts synonym relations with the global distributional features, while the pa ern module considers the local contexts mentioning both target strings. (3) Discover missing synonyms for the query entities with both the distributional module and the pa ern module.
3.1 Synonym Seed Collection
To automatically collect synonym seeds, our approach will rst detect entity mentions (strings that represent entities) in the given text corpus and link them to entities in the given knowledge base. A er that, we will retrieve the existing synonyms in knowledge bases as our training seeds. An illustrative example is presented in Figure 1.
Speci cally, we rst apply existing named-entity recognition (NER) tools [8]1 to detect entity mentions and phrases in the given text corpus. en some entity linking techniques such as the DBpedia Spotlight [3]2 are applied, which will map the detected entity mentions to the given knowledge base. During entity linking, some mentions can be linked to incorrect entities, leading to false synonym seeds. To remove such seeds, for each mention and its linked entity, if the surface string of that mention is not in the existing synonym list of that entity, we will remove the link between the mention and the entity,
A er entity mention detection and linking, the synonym seeds will be collected from the linked corpus. Speci cally, for each entity, we collect all mentions linked to that entity, and treat all corresponding surface strings as synonym seeds.
3.2 Methodology Overview
A er extracting synonym seeds from knowledge bases, we formulate an optimization framework to jointly learn the distributional module and the pa ern module.
To preserve the semantic meanings of di erent strings, our framework introduces a low-dimensional vector (a.k.a. embedding) to represent each entity surface string (i.e., strings that are linked to entities in knowledge bases) and each unlinkable string (i.e., words and phrases that are not linked to any entities). For the same strings that linked to di erent entities, as they have di erent semantic meanings, we introduce di erent embeddings for them. For example, the string "Washinton" can be linked to a state or a person, and we use two embeddings to represent Washinton (state) and Washinton (person) respectively.
e two modules of our framework are built on top of these string embeddings. Speci cally, both modules treat the embeddings as features for synonym prediction, and in turn update the embeddings based on the supervision from the synonym seeds, which may bring stronger predictive abilities to the learned embeddings. Meanwhile, since the string embeddings are shared between the two modules, each module is able to leverage the knowledge discovered by the other module, so that the two modules can mutually enhance to improve the learning process.
1 h p://stanfordnlp.github.io/CoreNLP/ 2 h ps://dbpedia-spotlight/dbpedia-spotlight
e overall objective of our framework is summarized as follows:
O = OD + OP,
(1)
where OD is the objective of the distributional module and OP is the objective of the pa ern module. Next, we introduce the details
of each module.
3.2.1 Distributional Module. e distributional module of our framework considers the global distributional features for synonym discovery. e module consists of an unsupervised part and a supervised part. In the unsupervised part, a co-occurrence network encoding the distributional information of strings will be constructed, and we try to preserve the distributional information into the string embeddings. Meanwhile in the supervised part, the synonym seeds will be used to learn a distributional score function, which takes string embeddings as features to predict whether two strings are synonymous or not. Unsupervised Part. In the unsupervised part, we rst construct a co-occurrence network between di erent strings, which captures their distributional information. Formally, all strings (i.e., entity surface strings and other unlinkable strings) within a sliding window of a certain size w in the text corpus are considered to be co-occurring with each other. e weight for each pair of strings in the co-occurrence network is de ned as their co-occurrence count.
A er network construction, we aim to preserve the encoded distributional information into the string embeddings, so that strings with similar semantic meanings will have similar embeddings. To preserve the distributional information, we observe that the cooccurrence counts of strings are related to the following factors.
O
3.1 (C
O
). (1) If two
strings have similar semantic meanings, then they are more likely to
co-occur with each other. (2) If a string tends to appear in the context
of another one, then they tend to co-occur frequently.
e above observation is quite intuitive. If two strings have simi-
lar semantic meanings, they are more likely to be mentioned in the
same topics, and therefore have a larger co-occurrence probability. For example, the strings "data mining" and "text mining" are highly
correlated, while they have quite di erent meanings from the word
"physics", and we can observe that the co-occurrence chances between "data mining" and "text mining" are much larger than those between "data mining" and "physics". On the other hand, some
string pairs with very di erent meanings may also have large co-
occurrence counts, when one tends to appear in the context of the
other one. For example, the word "capital" o en appears in the context of "USA", even they have very di erent meanings.
To exploit the above observation, for each string u, besides its embedding vector xu , we also introduce a context vector cu , which describes what kinds of strings are likely co-mentioned with u. Given a pair of strings (u, ), we model the conditional probability p(u| ) as follows:
p (u |
)
=
exp(xTu x + Z
xTu c
) ,
(2)
where Z is a normalization term. We see that if u and have
similar embedding vectors, meaning they have similar semantic meanings, the rst part (xTu x ) of the equation will be large, leading to a large conditional probability, which corresponds to the rst
observation 3.1. On the other hand, if the embedding vector of u is similar to the context vector of , meaning u tends to appear in the
context of , the second part (xTu c ) becomes large, which also leads to a large conditional probability, and this process corresponds to
the second observation 3.1.
To preserve the distributional information of strings, we expect the estimated distribution p(?| ) to be close to the empirical distribution p (?| ) (i.e., p (u | ) = wu, /d , where wu, is the cooccurrence count between u and , and d is the degree of in the network) for each string . erefore, we minimize the KL distance between p(?| ) and p (?| ), which is equivalent to the following objective [24]:
LC =
wu, log p(u | ),
(3)
u, V
where V is the vocabulary of all strings.
Directly optimizing the above objective is computational expen-
sive since it involves traversing all strings in the vocabulary when
computing the conditional probability. erefore, we leverage the
negative sampling techniques [9] to speed up the learning process, which modify the conditional probability p(u| ) in Eqn. 3 as follows:
N
log (xTu x + xTu c ) + Eun Pne (u)[1 - log (xTun x + xTun c )],
n=1
(4)
where (x) = 1/(1+exp(-x)) is the sigmoid function. e rst term
tries to maximize the probabilities of some observed string pairs,
while the second term tries to minimize the probabilities of N noisy pairs, and un is sampled from a noisy distribution Pne (u) du3/4 and du is the degree of string u in the network.
Supervised Part. e unsupervised part of the distributional mod-
ule can e ectively preserve the distributional information of strings
into the learned string embeddings. In the supervised part, we will
utilize the collected synonym seeds to train a distributional score
function, which treats the string embeddings as features to predict
whether two strings have the synonym relation or not.
To measure how likely two strings are synonymous, we intro-
duce a score for each pair of strings. Inspired by the existing
study [36], we use the following bilinear function to de ne the score of a string pair (u, ):
ScoreD (u, ) = xu WD xT ,
(5)
where xu is the embedding of string u, WD is a parameter matrix for the score function. Due to the e ciency issue, in this paper we
constrain WD as a diagonal matrix. To learn the parameters WD in the score function, we expect
that the synonymous string pairs could have larger scores than
those randomly sampled pairs. erefore we adopt the following
ranking based objective for learning:
LS =
min(1, ScoreD (u, ) - ScoreD (u, )), (6)
(u, )Sseed V
where Sseed is the set of synonymous string pairs, is a string randomly sampled from the string vocabulary. By maximizing the
above objective, the learned parameter matrix WD will be able to distinguish those synonymous pairs from others. Meanwhile, we
will update the string embeddings to maximize the objective, which
will bring more predictive abilities to the learned embeddings.
3.2.2 Pa ern Module. For a pair of target strings, the pa ern module of our framework predicts their relation from the sentences mentioning both of them. We achieve this by extracting a pa ern
Sentence Pattern Lexical Features Syntactic Features
Illinois , which is also called IL , is a state in the US . (ENT NNP nsubj) (called VBN acl:relcl) (ENT NN xcomp)
Embedding[called] NNP VBN NNP (NNP,VBN) (VBN,NNP) nsubj acl:relcl xcomp (nsubj,acl:relcl) (acl,xcomp)
Sentence Pattern Lexical Features Syntactic Features
Michigan , also known as MI , consists of two peninsulas . (ENT NNP nsubj) (known VBN acl) (ENT NNP xcomp)
Embedding[known] NNP VBN NNP (NNP,VBN) (VBN,NNP) nsubj acl xcomp (nsubj,acl) (acl,xcomp)
Figure 4: Examples of patterns and their features. For a pair of target strings (red ones) in each sentence, we de ne the pattern as the triples in the shortest dependency path. We collect both lexical features and syntactic features for pattern classi cation.
from each of such sentences, and collecting some lexical features and syntactic features to represent each pa ern. Based on the extracted features, a pa ern classi er is trained to predict whether a pa ern expresses the synonym relation between the target strings. Finally, we will integrate all prediction results from these pa erns to decide the relation of the target strings.
We rst introduce the de nition of the pa ern used in our framework. Following existing pa ern based approaches [11, 35], given two target strings in a sentence, we de ne the pa ern as the sequence of triples collected from the shortest dependency path connecting the two strings. Two examples can be found in Figure 4.
For each pa ern, we will extract some features and predict whether this pa ern expresses the synonym relation. We expect that the extracted features could well capture the functional correlations between pa erns. In other words, pa erns expressing synonym relations should have similar features. For example, consider the two sentences in Figure 4. e pa erns in both sentences express the synonym relation between the target strings (strings with the red color), and therefore we anticipate that the two pa erns could have similar features.
Towards this goal, we extract both lexical and syntactic features for each pa ern. For the lexical features, we average all embeddings of strings in a pa ern as the features. As the string embeddings can well preserve the semantic meanings of strings, such lexical features can e ectively capture the semantic correlations between di erent pa erns. Take the sentences in Figure 4 as an example. Since the strings "called" and "known" usually appear in similar contexts, they will have quite similar embeddings, and therefore the two pa erns will have similar lexical features, which is desirable. For the syntactic features, we expect that they can capture the syntactic structures of the pa erns. erefore for each pa ern, we treat all n-grams (1 n N ) in the part-of-speech tag sequence and the dependency label sequence as its syntactic features. Some example are presented in Figure 4, where we set N as 2.
Based on the extracted features, a pa ern classi er will be trained, which predicts whether a pa ern expresses the synonym relation. To collect positive examples for training, we extract pa erns from all sentences mentioning a pair of synonymous strings, and treat these pa erns as positive examples. For the negative examples, we randomly sample some string pairs without the synonym relation, and treat the corresponding pa erns as negative ones. We select the linear logistic classi er for classi cation. Given a pa ern pat
and its feature vector fpat , we de ne the probability that pa ern pat expresses the synonym relation as follows:
1
P ( pat = 1|pat ) = 1 + exp(-WPT fpat ) ,
(7)
where WP is the parameter vector of the classi er. We learn WP by maximizing the log-likelihood objective function, which is de ned
as below:
OP =
log P( pat |pat),
(8)
pat Spat
where Spat is the set of all training pa erns, pat is the label of pattern pat. By maximizing the above objective, the learned classi er
can e ectively predict whether a pa ern expresses the synonym re-
lation or not. Meanwhile, we will also update the string embeddings
during training, and therefore the learned string embeddings will
have be er predictive abilities for the synonym discovery problem.
A er learning the pa ern classi er, we can use it for synonym prediction. Speci cally, for a pair of target strings u and , we
rst collect all sentences mentioning both strings, and extract cor-
responding pa erns from them, then we measure the possibility
that u and are synonymous using the following score function ScoreP (u, ):
ScoreP (u, ) =
pat Spat (u,
) P(
pat = 1|pat ) ,
|Spat (u, )|
(9)
where Spat (u, ) is the set of all corresponding pa erns. Basically, our approach will classify all corresponding pa erns, and di erent pa erns will vote to decide whether u and are synonymous.
4 MODEL LEARNING AND INFERENCE
In this section, we introduce our optimization algorithm and how
we discover missing synonyms for entities.
Optimization Algorithm. e overall objective function of our framework consists of three parts. Two of them (LC and LS ) are from the distributional module and the other one (OP ) is from the pa ern module. To optimize the objective, we adopt the edge sam-
pling strategy [24]. In each iteration, we alternatively sample a
training example from the three parts, and then update the corre-
sponding parameters. We summarize the optimization algorithm
in Algorithm 1
Synonym Inference. To infer the synonyms of a query entity, our framework leverages both the distributional module and the
pa ern module. Formally, given a query entity e, suppose its name strings col-
lected from knowledge bases is Ss n (e). en for each candidate string u, we measure the possibility that u is a synonym of e using the following score function:
Score(e, u) =
{ScoreD (s, u) + ScoreP (s, u)}. (10)
s Ss n (e )
ScoreD (Eqn. 5) and ScoreP (Eqn. 9) are used to measure how likely two target strings are synonymous, which are learned from the distributional module and the pa ern module respectively. is a
parameter controlling the relative weights of the two parts. e
de nition of the score function is quite intuitive. For each candidate
string, we will compare it with all existing name strings of the query
entity, and these existing name strings will vote to decide whether
the candidate string is a synonym of the query entity.
................
................
In order to avoid copyright disputes, this page is only a partial summary.
To fulfill the demand for quickly locating and searching documents.
It is intelligent file search solution for home and business.
Related download
- what is practical knowledge christoph lumer university
- multi view embedding based synonyms for email search
- synonyms for words commonly used in resumes
- automatic synonym discovery with knowledge bases
- personal competencies dictionary attention to detail
- bloom s taxonomy of measurable verbs utica college
- building vocabulary knowledge by teaching paraphrasing
- knowledge management
- surfcon synonym discovery on privacy aware clinical data
Related searches
- words with synonym and antonym
- worked with synonym resume
- synonym for work well with others
- synonym for working well with others
- synonym for working with others
- synonym for working with people
- general knowledge trivia with answers
- general knowledge trivia with answer
- army bases with airborne units
- synonym for with that said
- exponents with integer bases calculator
- general knowledge quiz with answers